Agentic AI systems — software that takes actions, accesses tools, and completes tasks without human approval at each step — are already operating inside Singapore's supervisory framework for financial institutions. Firms have been running these systems with no new rules, no checklists, and no deadlines handed over. That is not reassurance. It is the clock, already running.
On October 7, the Monetary Authority of Singapore issued its final Guidelines on Artificial Intelligence Risk Management. The problem now has a due date. Two, in fact.
The agent chapter, however, is still being written.
What do the MAS AI guidelines require?
The MAS AI risk management guidelines apply to all financial institutions and all forms of AI. Each firm may decide how to meet them based on the nature and scale of its AI use and the associated risk materiality.
Four expectations carry the weight:
- Oversight: the board and senior management own AI risk, with clear roles, risk appetite, and frameworks.
- Life cycle: identify AI use, keep an inventory, assess materiality, and apply controls. These include data governance, testing, human oversight, cybersecurity, monitoring, and change management.
- Third parties: firms remain accountable for AI they buy in.
- Proportionality: basic policies are enough where failure would not have a material impact.
The timeline comes in two stages. Oversight and risk identification apply from October 7, 2027. Life cycle controls and capabilities follow by October 7, 2028.
MAS Deputy Managing Director Ho Hern Shin said that with greater regulatory clarity, "FIs can innovate with confidence."
Confidence, it turns out, arrives in instalments.
When is a basic AI policy enough?
This summer, the Financial Stability Board consulted on sound practices for responsible AI adoption. Michelle Bowman, who chairs the FSB's Standing Committee on Supervisory and Regulatory Cooperation, asked a pointed question. Does the framework strike the right balance between large and small institutions?
Several small banks answered politely. The case studies, they said, were hard to apply at their size.
MAS heard a version of the same concern from its own respondents. They wanted to know when basic policies would suffice, whether existing governance structures would do, and how to handle AI embedded in other services.
Singapore's answer is a test of impact, not novelty. If poor performance or unavailability of an AI tool is unlikely to materially affect the firm, its customers, or other institutions, basic policies are enough.
This is the right test. Autonomy is a dial, not a category. What matters is not how clever a tool is. It is what breaks when the tool fails.
What the Guidelines leave open is a harder question. The release speaks of benefits: better customer outcomes, productivity, new products. It does not say how anyone will measure them. Governance now has a deadline. Value still does not.
Are agents in scope before the guidance is published?
The most interesting sentence in the MAS AI risk management guidelines is easy to miss. MAS tells firms to review their controls as technology evolves, "such as the greater use of agentic AI systems that can operate autonomously and access tools." Then it adds: "In 2027, MAS intends to further consult the financial sector on what additional guidance on agentic AI would be useful."
Read that twice. Agents are in scope today. The detailed guidance is still a consultation away. And a consultation scheduled for 2027 is no defence for an incident in 2026.
The industry has not waited. In its FSB response, JPMorgan proposed an eight-part agentic governance framework, including agent identity, circuit breakers, and named human accountability. Visa argued that agentic AI is not inherently high risk and that agent identity and authority deserve attention.
So the drafts exist. The open question is whose draft becomes the template.
The agent rulebook will be written by whoever shows up. Last time, the vendors did. In the FSB consultation, vendors and consultancies outnumbered regulated institutions more than two to one. If banks want a say in what Singapore writes next, 2027 is the year to file in their own name.
Can firms use their existing governance structure?
Respondents to the MAS consultation asked whether they could use their existing governance structures. MAS said yes, as long as those structures provide adequate oversight and cross-functional coordination. Nobody needs a dedicated AI committee just to meet the expectation.
What adequate oversight looks like in practice is worth spelling out. The board sets risk appetite and owns accountability for AI outcomes. Senior management translates that appetite into frameworks, roles, and escalation paths. Cross-functional coordination means the people who build, buy, and monitor AI are not operating in separate silos — compliance, technology, and the business need a shared view of what is running and who owns each decision point.
The board and senior management still carry the responsibility. Existing structures just need to demonstrate that oversight, clearly and consistently.
Who owns the risk when AI comes from a third party?
The third-party section is where the Guidelines are firmest. Firms stay accountable for AI in the services they deliver, including AI developed, operated, or provided by others.
They must obtain sufficient assurance from providers, assess whether third-party AI suits its intended use, and apply compensating controls where assurance falls short. If the risk cannot be brought within appetite, they should consider limiting, suspending, or replacing the service.
This matters because, as Merchants Bancorp told the FSB, AI risk is now largely third-party risk.
So what does a compensating control look like? Deutsche Börse offered a useful answer in its FSB response: advance notice of model changes, and the right to pin a model version. A foundation model can change its behaviour with an update nobody announced. Contracts can fix that. Policies cannot.
"Replace the provider" sounds easy in a guideline and expensive in a contract.
Who can stop your AI — and have they ever tried?
Three questions frame the governance test for any agentic workflow: can it be fully automated, does it create a regulated outcome, and who owns it?
The Guidelines list human oversight and monitoring among the life cycle controls. For agents, MAS has not yet said what good looks like. Until it does, add a fourth question:
Who can stop this step, and when did they last try?
The answer is not a policy document. It is a test result. A named owner who has never halted anything provides the appearance of governance without the substance. FINMA's AI guidance points the same way, framing risk partly around autonomy. The more a system can do on its own, the more the off switch matters — and the more important it becomes that someone has actually used it.
The 2027 deadline is for the inventory. The deadline for knowing what your agents do was yesterday.
See where your AI governance stands against the MAS timelines.
Book a compliance walkthrough with the WealthAi team. We will map your agentic workflows against the October 2027 oversight requirements and the 2028 lifecycle controls — and answer the fourth question the guidelines have not yet asked for you.